AI-Generated Images Match Real Data in Surface Analysis

In industries ranging from aerospace to manufacturing, the quality of hard coatings like aluminum oxide (Al₂O₃) is critical for durability and performance, but assessing surface roughness often requires expensive, high-resolution imaging and large datasets. A new study demonstrates that synthetic images generated by artificial intelligence can effectively substitute for real experimental data in classification tasks, offering a cheaper and faster alternative. This approach could transform materials engineering by lowering barriers to AI adoption and streamlining quality control processes.

The researchers found that augmenting authentic datasets with AI-generated synthetic images yields test accuracies comparable to using only experimental images. In binary classification experiments comparing high and low roughness levels, replacing one class's real images with synthetic ones did not reduce accuracy on authentic test sets. For example, in ternary classification involving high, medium, and low roughness, configurations using synthetic images achieved performance similar to those using only authentic data, with one setup reaching 98% accuracy for high-roughness images and 100% for normal and low-roughness images. This indicates that synthetic images capture the structural features necessary for reliable classification without introducing biases that harm generalization.

Ology involved creating authentic images of Al₂O₃ samples using a laser scanning confocal microscope, categorizing them into three roughness levels based on surface roughness (Sa) values. Synthetic images were generated using Stable Diffusion XL, an AI model that takes authentic images as inputs to produce realistic counterparts with controlled variations. Classification models were trained using TensorFlow on Google Teachable Machine, with experiments designed to test substitution scenarios: in binary tasks, synthetic images replaced authentic ones for one roughness level, while in ternary tasks, various combinations were tested. Hyperparameters like epoch count, batch size, and learning rate were systematically varied to assess robustness, with a baseline of 100 epochs, batch size 16, and learning rate 0.001.

Analysis, as shown in Figures 1 and 2, revealed that synthetic images closely mimic the visual and structural characteristics of authentic samples across all roughness levels. For instance, high-roughness synthetic images replicated the stochastic morphology and peak-to-valley transitions seen in real scans, while low-roughness versions captured uniform textures. Hyperparameter optimization, detailed in Figure 3, showed that classification accuracy remained stable across different epoch numbers and batch sizes, but was sensitive to learning rate, with excessively high rates decreasing performance. This consistency across experiments confirms that synthetic data do not introduce unusual sensitivity, allowing for efficient training with fewer resources.

Of this research are significant for materials engineering and beyond, as it suggests that reliable image-based classification may not require large volumes of costly experimental data. By using synthetic images, laboratories and manufacturers can reduce dependence on ultra-high-resolution imaging equipment, lower costs, and accelerate model development. This could expand AI applications in quality assessment and industrial inspection, making advanced analysis more accessible. However, the study's limitations include its focus on a single material (Al₂O₃) and specific classification task, leaving generalizability to other materials and imaging conditions unverified. Future work should explore more diverse datasets and fine-grained tasks to fully assess the robustness of synthetic substitution in materials image analysis.